Figma puts an agent inside the design canvas — right on our thesis.

1. 01

   Figma's in-canvas Design agent enters beta — agentic editing on the live artifact

   Figma is now generating and remixing inside the canvas, component- and token-aware, feeding its design‑to‑code funnel. Closest competitive move of the week to our "edit the real thing" premise — on design files, not code.

2. 02

   Liveblocks v3.19.3 fixes disconnects while loading large Storage states

   The realtime engine under our canvas patches an unexpected-disconnect bug during large/extended streaming and large initial Storage loads, plus a Storage chunking speedup. Action needed: bump to 3.19.3 and regression-sweep heavy documents.

3. 03

   Neon ships an Agent Plan and per-tenant provisioning patterns

   Usage-based plan for high-volume programmatic provisioning, a neon-for-agent-platforms skill, and an MCP create_branch parentId fork. Directly relevant if we ever provision Postgres per generated app.

4. 04

   Lovable adds Workspace Skills — the "skills" primitive is now table stakes

   Reusable saved instructions Lovable auto-applies, landing the same week as a research-preview MCP server. Figma Make, Lovable, Claude Code all carry a skills primitive now — the differentiation moves to what it actually buys inside the surface.

Figma's new in-canvas agent (gradual beta from May 20) generates and remixes designs, makes bulk edits, swaps states across flows, populates content, and documents the design system — all while respecting existing components and tokens. It's distinct from the Figma Make "Skills" we covered last week: this is the design-file agent itself, operating directly on the live artifact rather than a prompt-to-prototype surface.

The strategic read: this is the most direct move of the window against the premise we own — editing the running thing, not a spec of it. Figma's version runs on design files and feeds its design-to-code funnel (agent → Dev Mode → MCP → coding agent). Our provisional moat holds — bidirectional editing on real React code is still further down the value chain than design-file synthesis — but the question is whether Figma's distribution and file gravity outrun that premise.

v3.19.3 (week 21) patches an unexpected-disconnect bug that fired during large or extended streaming responses — notably when a client loads a substantial initial Storage state. A separate backend chunking optimization speeds those initial loads. Our multiplayer canvas loads exactly these large Storage states, so this is a fix aimed straight at our usage pattern.

The action is a patch bump and a regression sweep — open a heavy document in a real multi-cursor session and confirm no drift. This follows last week's v3.19.2 Yjs keep-alive fix and the completed Storage v2 migration; the engine under the canvas is in active hardening.

Neon shipped (May 22) an Agent Plan — usage-based pricing for high-volume programmatic provisioning — plus a neon-for-agent-platforms skill with TypeScript samples for dual-org setups, per-tenant provisioning, and consumption-API queries. The MCP create_branch tool gained an optional parentId, so an agent can fork any branch, not just the project default.

We're on Supabase, so none of this is operational for us today. The relevance is shape-of-the-future: if we ever provision a database per user-generated app, this is the pattern — and Neon-under-Databricks doubling down on "Postgres for AI agents" is a standing competitive-watch item for Supabase.

Last week's Tunnel-in-Wrangler (press t for a public preview URL) now also works through the Cloudflare Vite plugin (May 18) — the path closer to our Vite-adjacent stack. Same-network demos, cross-device QA, and sharing a canvas branch with Daniel without a staging deploy get easier.

Also in window (May 21): VPC Network bindings (network_id: "cf1:network") to reach a private network from a Worker, and Artifacts namespace / repo / token management from the Wrangler CLI. Both are enterprise-leaning — low immediate relevance, worth filing.

Lovable shipped Workspace Skills (May 18) — reusable saved instructions it auto-applies, created in workspace settings, uploaded, or auto-saved from a repeated workflow. The same drop included a Lovable MCP server in research preview, TanStack Start as the new default, and a batch of connectors. With Figma Make and Claude Code already carrying the primitive, a "skills" surface is now the category baseline, not a differentiator.

After last week's billing earthquake (programmatic Claude moves to a metered credit pool June 15) and the Stainless acquisition, this window was calmer: a Project Glasswing update (May 22) and a KPMG strategic alliance (May 19) putting Claude across a 276k-person workforce. The KPMG deal is an enterprise-credibility proof-point for the bet we just made consolidating on Anthropic.

Watch item, date-unconfirmed: self-hosted sandboxes (running agent tool-execution on your own infra while the agent loop stays on Anthropic's) have been surfacing in engineering / docs content. If that ships as a real, dated capability, it's strategically significant for us — own the execution environment, keep the orchestration managed. Flagging for a dig once a firm source lands.